Software 2.0: We need to parse a bunch of different job ads. We'll have a rule engine, decide based on keywords what to return, do some filtering, maybe even semantic similarity to descriptions we know match with a certain position, and so on
Software 3.0: We need to parse a bunch of different job ads. Create a system prompt that says "You are a job description parser. Based on the user message, return a JSON structure with title, description, salary-range, company, position, experience-level" and etc, pass it the JSON schema of the structure you want and you have a parser that is slow, sometimes incorrect but (most likely) covers much broader range than your Software 2.0 parser.
Of course, this is wildly simplified and doesn't include everything, but that's the difference Karpathy is trying to highlight. Instead of programming those rules for the parser ourselves, you "program" the LLM via prompts to do that thing.
